TensorFlow 2.6, the deep learning framework, is now past the beta testing phase and is moving to general availability. However, machine learning users should be careful since the release has a few breaking changes worth considering before switching over.
Among other things, version 2.6 finally delivers on the promise to get the TensorFlow implementation of Keras its own repository.
Before now, the project was not part of the Google-backed TensorFlow organization at all. It existed as an independent library that could be used with a CNTK and Theano backend, as well as TF.
As time went by, TensorFlow implementation superseded the original due to the addition of crucial features like eager section, distribution, and TPU training.
The Keras team decided to discontinue multi-backend Keras and focus on tf-keras. Keras creator, Francois Chollet, mentioned last year in June when Keras 2.4 came out that development under the TensorFlow banners might be a temporary option, calling keras-team/keras, the potential future home.
TensorFlow 2.6 realizes what Chollet said as it has gained separate PIP package status again. API endpoints will stay as they are, except they now will be backed by the Keras PIP package. Users importing tensorflow.python.keras have to figure out how to use the public tf.keras API.
Some of the significant changes made by the Keras team include the replacement of methods Model.to_yaml() and keras.models.model_from-yaml to raise a RuntimeError, since they can be used to cause arbitrary code execution.
Recommended alternatives include JSON and H5 serialization.
There are also stabilized pre-processing layers and relevant APIs, as well as a sidecar evaluator for programs (still experimental) designed for use on evaluator tasks or single-worker training.
To find out more information about what else is new in this release, read more here.